File size: 3,023 Bytes
f9b94bf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | """Hugging Face configuration for MiniCPM-RobotTrack."""
from typing import Any, Dict, Optional
from transformers.configuration_utils import PretrainedConfig
class MiniCPMRobotTrackConfig(PretrainedConfig):
"""Configuration for the MiniCPM-RobotTrack trajectory policy."""
model_type = "minicpm_robottrack"
has_no_defaults_at_init = True
def __init__(
self,
backbone_config: Optional[Dict[str, Any]] = None,
vision_feature_dim: int = 1536,
history_frames: int = 31,
coarse_tokens_per_frame: int = 4,
fine_tokens_current_frame: int = 64,
num_waypoints: int = 8,
action_dim: int = 3,
max_text_tokens: int = 128,
max_time_steps: int = 4096,
trajectory_dropout: float = 0.4,
xy_scale: float = 2.0,
use_tanh_actions: bool = True,
backbone_dtype: str = "bfloat16",
**kwargs: Any,
) -> None:
self.backbone_config = dict(backbone_config or {})
self.vision_feature_dim = int(vision_feature_dim)
self.history_frames = int(history_frames)
self.coarse_tokens_per_frame = int(coarse_tokens_per_frame)
self.fine_tokens_current_frame = int(fine_tokens_current_frame)
self.num_waypoints = int(num_waypoints)
self.action_dim = int(action_dim)
self.max_text_tokens = int(max_text_tokens)
self.max_time_steps = int(max_time_steps)
self.trajectory_dropout = float(trajectory_dropout)
self.xy_scale = float(xy_scale)
self.use_tanh_actions = bool(use_tanh_actions)
self.backbone_dtype = str(backbone_dtype)
self._validate_robot_track_config()
super().__init__(**kwargs)
def _validate_robot_track_config(self) -> None:
if not self.backbone_config:
raise ValueError("backbone_config must contain the bundled MiniCPM configuration")
if self.vision_feature_dim <= 0:
raise ValueError("vision_feature_dim must be positive")
if self.history_frames < 0:
raise ValueError("history_frames cannot be negative")
for name, value in (
("coarse_tokens_per_frame", self.coarse_tokens_per_frame),
("fine_tokens_current_frame", self.fine_tokens_current_frame),
):
side = int(round(value**0.5)) if value > 0 else 0
if side * side != value:
raise ValueError(f"{name} must be a positive square number")
if self.num_waypoints < 2:
raise ValueError("num_waypoints must be at least 2")
if self.action_dim != 3:
raise ValueError("actions must use [x, y, yaw] format")
if self.max_text_tokens <= 0 or self.max_time_steps <= 0:
raise ValueError("token and time limits must be positive")
if not 0.0 <= self.trajectory_dropout < 1.0:
raise ValueError("trajectory_dropout must be in [0, 1)")
if self.xy_scale <= 0.0:
raise ValueError("xy_scale must be positive")
|